The Application Of Recurrent Neural Network On Image Super Resolution

High-resolution images contain rich details and sensitive colors.In many applications,such as video monitoring,medical diagnosis and text recognition,high resolution images are urgently needed to improve the effect and accuracy of subsequent processing.Image super-resolution reconstruction technology refers to the process of reconstructing one or more high-resolution images from a low-resolution image or image sequence.The technology provides a way to improve image resolution without increasing the cost of hardware.In recent years,the image super-resolution based on learning has attracted many scholars' attention,and its core idea is to establish a nonlinear mapping of low-resolution images to high-resolution images.In the process of data processing,the input and output relationship of the recurrent neural network is connected by the function of activation function,which can be regarded as a highly nonlinear mapping relation.In addition,the particularity of the recurrent neural network structure enables it to remember and use the input data in the previous time for the current input,so that it has more efficient parallel processing ability and learning ability.In this paper,an image super-resolution algorithm based on recurrent neural network is studied.The main tasks are as follows:Firstly,an image super-resolution reconstruction algorithm based on(Recurrent Neural Network,RNN)is proposed.In this algorithm,recurrent neural network is used as a mapping model to obtain a nonlinear mapping relationship between high and low resolution images through autonomous learning of the network.A large number of high and low resolution grayscale images are taken directly for block processing and normalization.Then,the network weight is updated through the training phase to obtain the ideal network structure.Finally,the training completed network is reconstructed to obtain high resolution images.Experimental results show that this algorithm has better edge effect and rich texture details.Compared with other algorithms,this algorithm also has higher PSNR value and shorter time consumption.Secondly,the number of layers in the time dimension is too deep in the training process of RNN,which can easily lead to explosion or gradient disappeared.Besides,learning process can't be relied on for a long time and the image quality is adversely affected.Therefore,an image super-resolution reconstruction algorithm based on(Long Short Term Memory,LSTM)network is proposed.The algorithm use LSTM network as a new mapping model for learning,which use the door control unit to improve the structure of RNN,can effectively avoid gradient disappeared in the process of network training.Finally,due to the complex structure of LSTM network model,there is a long time consumption in training process and image reconstruction process.Therefore,this paper presents an image super-resolution reconstruction algorithm based on(Gated Recurrent Units,GRU)network.GRU as a kind of improved network structure of LSTM,reduce the number of the door control unit.Which both to solve the problem of gradient disappeared,and can improve the time-consuming problem of LSTM network to improve the efficiency of network operation.The experimental results show that the high-resolution images recovered through GRU network model not only have higher PSNR value and better visual effect,but also significantly improve the running speed.